Background: Artificial intelligence (AI) applications in dentistry are advancing rapidly, yet external validation studies are scarce. This pilot study evaluated 2 US Food and Drug Administration-cleared AI-based decision support systems for detecting caries and clinically relevant periodontal bone loss (≥ 33% or ≥ 5 mm of attachment loss) using longitudinal clinical data to establish ground truth.
Methods: A retrospective analysis of 90 patients (mean [SD] age, 61.4 [13.2] years) records was conducted at the Veterans Affairs Greater Los Angeles Health Care System in Los Angeles, California. Each case included full-mouth radiographs and standardized clinical documentation with 6 through 12 months of follow-up. Ground truth was determined through a multiphase validation protocol involving calibrated examiners and adjudication. Analyses emphasized specificity and negative predictive value (NPV) to assess diagnostic concordance.
Results: For caries detection, vendor A achieved 78.68% concordance, 80.17% specificity, and 96.92% NPV; vendor B achieved 82.43%, 83.76%, and 97.36%, respectively. For periodontal bone loss, vendor A achieved 78.47% concordance, 94.23% specificity, and 80.64% NPV; vendor B achieved 73.89%, 93.55%, and 75.20%, respectively.
Conclusions: The results of this external validation pilot study showed AI tools can reliably exclude disease and reduce false-positive results. The use of retrospective longitudinal ground truth with standardized validation protocols provides a practical template for future multisite studies.
Practical implications: AI-based decision support tools are most effective as adjunctive screening aids in dental practice. Their high specificity and NPV allow clinicians to trust negative findings, avoid unnecessary interventions, and maintain clinical oversight of positive findings requiring confirmation.
扫码关注我们
求助内容:
应助结果提醒方式:
